maria-discuss team mailing list archive

[Kristian Nielsen <knielsen@xxxxxxxxxxxxxxx>]

Laurynas Biveinis <laurynas.biveinis@xxxxxxxxx> writes:

> We have another platform-specific addition: thread-local storage
> implemented by __thread GCC keyword, which is GNU specific. This is

Hm, this is interesting, I was not aware of the __thread feature. This looks
to be much more efficient than pthread_getspecific(), which is one of the top
functions in the profiles for sysbench readonly:

    static __thread long tcount = 0;

    long
    bump(long del)
    {
      return (tcount += del);
    }


    0000000000000000 <bump>:
       0:       48 89 f8                mov    %rdi,%rax
       3:       64 48 03 04 25 00 00    add    %fs:0x0,%rax
       a:       00 00 
       c:       64 48 89 04 25 00 00    mov    %rax,%fs:0x0
      13:       00 00 
      15:       c3                      retq   

It is just a single instruction using %fs: addressing to access the
thread_local storage. Apparently the linker resolves the %fs: offset in the
resulting binary:

    0000000000400580 <bump>:
      400580:       48 89 f8                mov    %rdi,%rax
      400583:       64 48 03 04 25 e0 ff    add    %fs:0xffffffffffffffe0,%rax
      40058a:       ff ff 
      40058c:       64 48 89 04 25 e0 ff    mov    %rax,%fs:0xffffffffffffffe0
      400593:       ff ff 
      400595:       c3                      retq   

    00000000004005a0 <bump2>:
      4005a0:       48 89 f8                mov    %rdi,%rax
      4005a3:       64 48 03 04 25 f0 ff    add    %fs:0xfffffffffffffff0,%rax
      4005aa:       ff ff 
      4005ac:       64 48 89 04 25 f0 ff    mov    %rax,%fs:0xfffffffffffffff0
      4005b3:       ff ff 
      4005b5:       c3                      retq   

So that could be very promising for an easy way to tweak out a little more
performance. Something to look into at some point.

 - Kristian.